Machine Learning Papers

OWL 2 DL ontologies, grounded in the description logic $\mathcal{SROIQ}$, express large knowledge bases in biomedicine and the Semantic Web. Neuro-symbolic (NeSy) learners over description logics either embed the ontology in a continuous space, abandoning classical entailment, or restrict to the Horn fragment $\mathcal{EL}^{++}$, which has a single canonical model. We present Baobab, which compiles a $\mathcal{SROIQ}$ ontology with a finite ABox into a Sentential Decision Diagram (SDD): it saturates a propositional core under a consequence-based calculus and instantiates the remaining $\mathcal{SROIQ}$ features (nominals, number restrictions, and the role axioms) over the active domain. The SDD's evidence-conditioned weighted model count then trains a perception network to recognize real images under partial ABox supervision: on an ontology that exercises every distinctive $\mathcal{SROIQ}$ feature, a CNN learns to read MNIST digits coupled by a successor relation and recovers latent ontology concepts that an independent perception leaves at chance. When the supervision admits several ontology-consistent completions, an independent perception collapses onto one, a reasoning shortcut: we show that a mixture indexed by the query's justifications can represent the calibrated posterior no independent perception can, and that seeding it from the circuit's enumerated completions attains the Bayes-optimal posterior on a real-image MNIST task where single-WMC and learned mixtures (the BEARS-ensemble hypothesis class) do not: to our knowledge the first to characterize and mitigate reasoning shortcuts in a non-Horn description logic. Soundness of the compiler and the representation result are machine-checked in Lean 4. Code is available at https://github.com/bio-ontology-research-group/baobab.

Browser agents perform well on short, clean demonstrations, but real deployment is fundamentally different: agents must sustain dozens of decisions on live websites while recovering from mistakes and navigating complex UIs. We argue that closing this gap requires alignment at every level of the pipeline, including execution, supervision, optimization, and evaluation, rather than scale alone. We present Wuying-Browser-Agent, a unified framework that addresses each of these levels. A structured browser harness provides stable execution primitives and decision-oriented context management. Reflection and UI-specialized Curriculum SFT (RUIC-SFT) explicitly trains on recovery trajectories and complex-UI interactions. Divergence-Aware Online GRPO (DAO-GRPO) improves long-horizon credit assignment through potential-based reward shaping and divergence-aware step weighting. Finally, we introduce BrowserBench, a bilingual real-web benchmark of 350 tasks averaging 37.9 steps, because most existing benchmarks are too short to expose long-horizon failure modes. Wuying-Browser-Agent-27B achieves 80.6\% on WebVoyager, 66.7\% on Online-Mind2Web, and 65.1\% on BrowserBench, establishing a new open-source state of the art on browser-use benchmarks. The same pipeline also transfers beyond browser use, demonstrating strong general agentic ability and reaching an average score of 73.8 on Tau2-Bench, Claw-Eval, and BFCL-v4.

Reliable optimization is central to neural network (NN) training, yet Adam, the default optimizer for modern LLMs, rests on a fragile foundation. This thesis develops a principled grounding for Adam and motivates new designs. First, we revisit Adam's divergence--convergence debate and show the existence of a problem-dependent phase transition: with properly chosen, batch-size-dependent hyperparameters, Adam converges, whereas under small-$β_2$ regimes it can diverge. Second, we investigate why Adam substantially outperforms SGD on Transformers through Hessian structure. We find that the Hessian evolves toward a near-block-diagonal form along training, accompanied by strong block heterogeneity. We prove that this structure makes Adam's diagonal preconditioner effective. We further show that this special Hessian structure originates from consecutive multiplications of large matrix variables, and we provide a rigorous analysis based on random matrix theory. Finally, these insights motivate Adam-mini, a new optimizer that reduces Adam's memory footprint by 50\% while preserving its performance. Our results also have broader implications beyond Adam: they reveal new local structures in matrix-based nonconvex problems, and also help understand and improve recent NN optimizers, such as Muon.

Retrieval-augmented generation (RAG) answers a question by retrieving passages from a vector store and trusting them as context, so anyone who can add documents can try to steer the answer. A recent, appealing defense filters poisoning at ingestion, rejecting any document that behaves like a hub. We show it -- and every ingestion-time filter -- is defeated by a coordinated adversary that injects a handful of individually unremarkable documents which together surround one target query and seize its top-k (on BGE-large / BEIR, m=10 documents take 10/10; 9.9/10 on a live HNSW index). The attack is not theoretical. Realized as ordinary fluent text and run end-to-end through a BGE-large + HNSW + Qwen2.5-7B pipeline, it makes the generator emit the attacker's planted claim in 88% of targets, versus 0% without the injection. And no admission-time defense stops it: at ingestion an attack cone is geometrically identical to a legitimate niche upload, so -- measuring this directly -- the strongest trained classifier, given every feature and thousands of examples, separates the two no better than chance, catching 4.2% of attacks at a 1% false-positive rate. We prove this limit for the entire class of ingestion-time statistics (any decision from documents and reference queries alone), and it reproduces -- and worsens -- across two corpora and five encoders. The one signal that separates an attack from legitimate niche ingestion -- a query's demand -- is invisible before retrieval, which is also the escape: a retrieval-time detector that observes demand catches 100% of the attacks at the same 1% false-positive rate. Coverage of the query space by an admission gate is not containment of coordinated poisoning; robust defense must move past the front door, to demand.

Most LLM-based automated algorithm design methods optimize a designated component within a human-specified scaffold, fixing overall organization and component interactions. We present ATLAS, an embedding-guided quality-diversity framework for scaffold-free full-algorithm synthesis in combinatorial optimization. The problem specification supplies objectives and constraints; a minimal I/O interface fixes only instance and solution formats; the LLM chooses and restructures components, interactions, and control flow. This freedom enlarges the search space, risking invalid candidates and premature convergence to one design region. ATLAS independently detects execution, interface, and feasibility failures, recomputes objectives, and applies error-conditioned repair; similarity-based archive management preserves algorithms across embedding-space regions to counter premature convergence. Its three-layer search refines the best design, gives other regions dedicated refinement opportunities, and performs cross-region synthesis to recombine components and their interactions. Across four NP-hard problems, ATLAS outperforms several state-of-the-art component-synthesis methods and a matched full-synthesis baseline while remaining competitive with strong human-designed algorithms. One ATLAS run retains several algorithms with comparable performance from distinct embedding-space regions rather than a single design. Code inspection finds that these multi-component designs differ in their primary construction or global-search backbone. Our results suggest that embedding-guided quality-diversity search can make the enlarged full-algorithm design space practically searchable. Source code and exact executable prompts are available at .

The growing ecosystem of large language models (LLMs) offers huge potential to optimize performance-cost trade-offs. However, their heterogeneous capabilities and inference costs make efficiently routing queries a significant challenge. Existing paradigms are inflexible: one-shot routers commit before observing responses, whereas conventional cascades stop adaptively but follow a fixed model order. Cascade routing removes both restrictions by reconsidering whether to stop or invoke another model after each response. Current methods use a predict-then-optimize pipeline estimating response quality and future model utility. However, prediction loss for quality or utility is not equivalent to routing-decision loss. A lower prediction error does not necessarily yield a better action; a small boundary-crossing error can reverse a ``stop'' or model-selection decision. Therefore, we propose RLCascadeRouter, a quality-estimator-free framework that formulates cascade routing as a Markov decision process with actions comprising ``stop'' and model selection. It uses trajectory returns and advantages to directly optimize the performance-cost objective. Its Cascade Policy Network models candidate complementarity for model selection and remaining-action value for stopping, eliminating independent post-hoc response-quality estimators. Evaluated across ten LLMRouterBench benchmarks with thirteen LLMs, RLCascadeRouter outperforms strong baselines and achieves superior performance-cost trade-offs. It incorporates unseen models without retraining, and ablation studies validate both policy components.

Agentic applications are shifting AI serving from isolated model inference to long-running workloads in which LLMs coordinate tools, environments, and persistent state. However, the system behavior of these workloads---where latency, cost, and bottlenecks arise---remains poorly characterized, leaving serving systems to rely on assumptions built for conventional inference. We present AgentSysBench, a benchmark suite and measurement toolkit with ten representative agentic applications and unified systems-level instrumentation. Across controlled deployments and production traces, we identify six properties that distinguish agentic workloads from conventional LLM serving: (1) execution is heavyweight and stateful, with non-LLM components dominating latency in 5 of 10 applications and sandbox working-set memory peaking at 28 GB per session; (2) applications compose components with heterogeneous resource affinity---GPU-bound inference, memory-bound retrieval, CPU-bound sandboxes---whose task latencies diverge by up to 32x; (3) bottlenecks shift across requests, models, and deployments; (4) production sessions hold state idle for minutes to hours between active steps; (5) a control-plane tax---auxiliary LLM calls and context overhead from tool schemas and observations---crowds out productive compute and context; and (6) production traces from three applications reveal heavy cross-request redundancy in search queries and web fetches, exposing a large caching opportunity. Four design explorations demonstrate that these findings are actionable: task-aware serving reduces latency by 29--40%, communication-aware placement by up to 4.5x, state offloading reduces memory usage by 4.6x, and tool-result caching removes 35.2% of redundant search calls and saves 19.3% of aggregate search latency.

In formal verification, both the autoformalization of statements and automated proof search have been studied extensively. While automated proof search can produce a formal proof that compiles, the generated proof does not necessarily reflect how the natural-language argument arrives at its conclusion--a property we refer to as faithfulness. With faithfully formalized proofs, one can check the reasoning behind a human- or AI-written argument, and assist mathematicians in formalizing their proof sketches. However, it is particularly challenging due to misalignment of formal proof tactics and natural language reasoning. In this work, we rigorously describe a set of five necessary conditions a faithful formal proof must satisfy, and introduce Pistis, an agentic, oracle-guided proof search that produces formal Lean proofs that satisfy them. At its core is a novel faithfulness-preserving divide-and-conquer search, which we name OrderDecompose, that tracks citation dependencies and blocks unfaithful shortcuts, paired with a refutation search, that surfaces gaps and errors in the natural language proof source. OrderDecompose completes proofs that baselines cannot close even within a 12-hour budget, and its artifacts compile over 33$\times$ as fast as prior work's. We apply Pistis on the first three books of Euclid's Elements, producing high-quality artifacts containing faithful formal proofs. Under a blinded human study and an LLM-as-a-judge protocol on rigorous rubrics, Pistis-generated proofs are favored over prior works--2.89$\times$ and 5.2$\times$ as often by human reviewers and the LLM judge, respectively. It further uncovers gaps in Euclid's proofs and their translation, and can accept or refute natural language proofs written by humans or AI, demonstrating that faithful formalization is useful as a proof-checking tool.

When a user question is underspecified, a capable model should recognize that its context is insufficient, identify the missing information, ask for it, and respond only once that information determines a unique answer. We formalize multi-turn information seeking as solving a k-underspecified constraint satisfaction problem, where k is the number of variables jointly required to determine the target and therefore measures the degree of missing information. We instantiate the formulation in MT-InfoSeek, a controlled evaluation suite of 5,251 problems and 9,006 task instances spanning mathematics, logic, biology, medicine, and general knowledge. We evaluate models along three axes: what they ask, when they ask it, and how the acquired information affects the final answer. Performance degrades across models and domains as underspecification increases. Models recognize that additional information is needed but underestimate how much, and in logical problems at k = 2 they under-predict the degree of missing information about four times as often as they over-predict it. They also fail to identify a minimal sufficient set of queries, improve only marginally when given the true k, and often stop before acquiring sufficient information. In tasks with ordered dependencies, an incorrect query order reduces final accuracy even when the model eventually acquires all necessary information. We measure information seeking directly through final sufficiency, which records whether the acquired information determines the target independent of answer generation. This separation shows differences between models that final accuracy alone does not capture, and indicates that the ability to seek information over multiple turns is distinct from the ability to generate answers and is not measured by current LLM evaluations.

Frontier LLM agents increasingly transact on behalf of separate principals, often using natural language rather than structured APIs. Much of the safety literature studies misaligned LLM behavior through adversarial-elicitation evaluations on single agents or stylized tasks. Its prevalence and structure in settings that combine long horizons, separate principals, real operational state, and inter-agent natural-language exchange remain insufficiently measured. We study 2,583 inter-agent emails from 20 one-year simulation runs of Vending-Bench Arena, a competitive vending environment spanning 13 frontier LLMs. We operationalize speech-act misalignment as emails containing false factual claims, manipulation, collusion, or threats, combining message content with ground-truth simulator state and logged reasoning traces to classify and validate such behavior. Under our primary classifier, 12.6% of emails are labeled misaligned; misalignment appears in all 20 runs and 74.7% of individual agent-runs. Both the magnitude and composition of this misalignment are preserved under repeated classification at different sampling temperatures and under full-pipeline replication with judges from two other frontier-model families. Misalignment is also reciprocal and stress-conditioned: receiving a misaligned email from a counterparty raises the odds of a misaligned reply by 1.65x, and low-inventory conditions raise them by 1.58x. Across tests of capability-asymmetric exploitation, we find no evidence that higher-capability models differentially exploit weaker counterparties, and model performance rank does not predict misalignment rates. Together, these results indicate that measurable, state-dependent misalignment can arise in competitive multi-agent environments without engineered elicitation, in patterns associated with operational scarcity and counterparty behavior rather than model capability alone.

We revisit the problem of learning predictors robust to adversarial examples at test-time. We prove that VC classes are adversarially robustly learnable with sample complexity linear in the VC dimension $d$, providing an exponential improvement over the previous upper bound of Montasser, Hanneke, and Srebro (2019). Remarkably, this result is achieved with a simple improper algorithm that combines the classic heuristic bagging (bootstrap aggregation) of Breiman (1996) with robust empirical risk minimization (RERM). Our algorithm computes RERMs on $O(d^\star)$ independent bootstrap samples and outputs their majority vote, where $d^\star$ denotes the dual VC dimension. We complement this result with a lower bound showing that this is unavoidable: in general, any learner in this oracle model requires $Ω(d^\star)$ calls to an RERM oracle, even when given arbitrarily many training examples.

Vision encoders are a critical component of vision-language models, and scaling their capacity effectively improves performance. However, dense scaling increases compute cost and inference latency. Mixture-of-Experts (MoE) architectures offer a compelling alternative, having enabled efficient scaling in LLMs, yet the MoE design space for CLIP-style vision encoders remains underexplored at State-of-the-Art (SOTA) levels. In this work, we systematically study MoE designs for vision encoder scaling and find that fine-grained MoE topologies yield substantial gains over both dense and standard MoE counterparts. We further propose an auxiliary-loss-free balancing variant for better expert utilization, and design a specialized MoE kernel to mitigate inference latency overhead. To enhance video capabilities while preserving image knowledge, we introduce frame-level distillation paired with a novel freezing mechanism. We pretrain a series of Mixture-of-Experts Vision Encoders (MoE-ViE) across a range of sizes, all consistently outperforming their dense counterparts. Our largest model matches the zero-shot performance of a SOTA encoder 1.7x its size at 76% of its latency. When aligned with an LLM, MoE-ViE surpasses all compared encoders on image and video benchmarks, including those with up to 5x more activated parameters. Code is available at https://github.com/facebookresearch/moe_vie.

Chain-of-thought reasoning has substantially improved the problem-solving capabilities of multimodal large language models. Fine-grained visual evidence, however, remains difficult to preserve and reuse across text-based reasoning steps. To address this limitation, tool-augmented thinking-with-images methods maintain visual access externally by revisiting or manipulating the image, but require predefined tools and additional inference-time processing. As an internal alternative, continuous visual latent reasoning retains intermediate computation in hidden states. However, its prevailing autoregressive construction makes each latent state depend on its predecessors, so later states may repeat information already present in the latent sequence rather than capture complementary visual details. We introduce GLaQ, a grounded latent-query framework that replaces sequential latent rollout with a fixed set of context-conditioned queries grounded in the original visual tokens. The grounded queries are reinjected for answer generation, providing direct and coordinated access to source visual evidence. We train GLaQ with localized-view supervision followed by reinforcement learning under task-level rewards. Across five benchmarks for fine-grained visual understanding and perception, GLaQ-7B gains 5.99--9.66\% over its base model and leads all compared visual latent methods, suggesting that direct query-to-image grounding can recover localized evidence from the full image without external visual operations or autoregressive latent rollouts.

Maintaining global geometric consistency is a central challenge in long-sequence 3D reconstruction, with scale drift being the most critical failure mode. In chunk-based inference pipelines, the scale degree of freedom in sequential Sim(3) alignment is left unconstrained, causing estimation errors to compound multiplicatively and distort global trajectories and point cloud geometry. We present a scale-consistency enhancement framework built on a key insight: in structured environments such as driving scenes, geometric quantities arising from environmental regularity remain inherently invariant across temporal segments, and discrepancies in their per-chunk measurements directly expose inter-chunk scale drift. We propose Scene Geometric Invariant Anchoring (SGIA), which extracts dominant geometric invariants from each chunk's predicted point cloud via coarse-to-fine robust estimation and exploits their cross-chunk consistency to establish scale constraints independent of point cloud registration, explicitly degenerating 7-DoF Sim(3) alignment into 6-DoF rigid-body transformation and severing chain-wise scale error propagation at its source. We further introduce a lightweight test-time adaptation strategy that fine-tunes only normalization-layer parameters via multi-objective self-supervision, progressively improving intra-chunk predictions along the sequence. Both modules are plug-and-play and require no offline retraining. Experiments on multiple long-sequence benchmarks demonstrate state-of-the-art performance, reducing absolute trajectory error by up to 32% with significant gains in trajectory stability and reconstruction quality. Code: https://github.com/WZ-CS/VGGT-Align

Interactive game world models typically autoregress visual observations directly in pixel or latent space, forcing structured properties such as pose, geometry, and occlusion to be implicitly maintained by the same generative sequence. Over long horizons, errors in these latent world properties accumulate, making consistency and controllability fragile. We explicitly model the evolving world state, delegate exact geometric computation to a fixed, zero-parameter renderer, and leave the neural model to synthesize appearance. We instantiate this idea as Marionette, a world model for interactive games with articulated characters. First, a two-stage autoregressive dynamics model predicts an explicit and interpretable 276-dimensional 3D world state comprising multi-entity articulated skeletons, metric root trajectories, and rotations. Second, a zero-parameter graphics bridge converts the predicted state into pose-control videos, computing world-space geometry and occlusion in closed form. Third, a control-conditioned video-diffusion observation model synthesizes photorealistic RGB observations from the resulting structured controls. Our experiments establish two properties of Marionette. First, the predicted world state is directly controllable. Forcing a mismatched action stream changes root-aligned joint error by 31% across 48 held-out segments. Second, long-horizon behaviour is determined in the state, and can be repaired there. Left free, the two generated characters drift to 21.2 m apart (recorded sessions stay near 5 m) and a third of frames show ground penetration. Two rules imposed on the explicit state, a terrain collider and a separation cap, cut penetration by 66% and keep the pair engaged, with no change to the observation model. Routing appearance through the predicted state costs no fidelity we can detect, at an FVD of 831 against 799 for recorded pose.

When visual evidence is occluded or chaotic, models should abstain. In this paper, we show that Vision-Language Models (VLMs) can internally distinguish when abstention is required, but fail to express it anyway. We introduce TRAPSBench, a procedurally generated video benchmark of 1,404 matched physics pairs in which a single targeted change renders the outcome undeterminable from the visual evidence. Furthermore, we introduce Penalized Epistemic Calibration Score (PECS), a new robust metric that requires models to both answer correctly when the outcome is knowable, and abstain when the outcome is not. Across 16 VLMs spanning five families, spontaneous restraint is poor: the best PECS is 0.292. The bottleneck is expression, not perception: linear probes decode answerability from hidden states at up to 0.91 AUROC across physics domains; steering a single-layer void direction causally induces or suppresses abstention. Our results replicate across three open-weight families (Qwen, Gemma, LLaVA). The failure is also more pronounced in visual than textual uncertainty: models detect textual impossibility about 4x more readily than missing visual evidence. Closing this representation--output gap likely requires output-stage interventions.

While 3D Vision-Language Models (3D VLMs) have demonstrated remarkable spatial reasoning capabilities, they suffer from massive visual token counts that create severe computational bottlenecks during inference. Existing token pruning methods primarily rely on diversity-based selection, discarding similar tokens to maximize dispersion. However, in 3D environments, this approach frequently drops representative prototype tokens in favor of outliers, breaking the multi-view consistencies and geometric structures essential for spatial reasoning. In this paper, we propose a paradigm shift for 3D VLM token pruning: from maximizing diversity to preserving visual evidence coverage. We introduce CoverPrune, a training-free framework that formulates inference-time token pruning as an Optimal Transport (OT) problem. To overcome the intractable combinatorial subset selection inherent in this formulation, we design the Feature-Spatial-Temporal (FST) transport cost and target capacity, along with an efficient Spatial-Guided Greedy Selection (SGS) algorithm to approximate the OT objective. Furthermore, we propose CoverPrune-Lite, an accelerated variant utilizing spatially structured local matching for minimal overhead. Extensive experiments across multiple 3D visual-spatial reasoning benchmarks demonstrate that our methods achieve state-of-the-art token efficiency, maintaining robust reasoning performance even under highly aggressive pruning budgets. Visit our project website at https://github.com/Brucess/CoverPrune.

Modern multimodal foundation models (MFMs) have made rapid progress on tasks requiring integrated perception across speech, vision, and language, including emotion recognition. However, it remains unclear whether they recognize speech and facial emotion through shared affective functional units or modality-specific pathways. We explore emotion-sensitive neurons (ESNs), sparse decoder neurons selectively associated with emotion categories, in three MFMs: Gemma-4-12B-it, MiniCPM-o-4.5, and Qwen2.5-Omni-7B. Using speech emotion recognition and facial expression recognition as complementary probes, we identify acoustic and visual ESNs. Visual ESNs are causally meaningful: deactivating them selectively impairs recognition of the associated facial emotion, whereas steering their activations selectively enhances recognition of that emotion relative to other emotion categories. Acoustic and visual ESNs further show emotion-matched overlap and similar layer-wise distributions, indicating partial structural alignment between affective representations across speech and faces. Finally, cross-modal interventions reveal bidirectional causal transfer: ESNs identified from one modality produce emotion-specific effects when applied to the other. Our findings provide one of the first cross-modality activation-level analyses of affective functional units in MFMs, suggesting that speech and facial emotion recognition partially converge onto sparse decoder-level components that can be localized and manipulated without training.

Stateful language agents assume a rejected branch can be taken back by clearing it from the application transcript. We show this breaks when the serving session retains key/value (KV) state across the logical abort: the model can continue attending to content the application believes it discarded. We formalize the missing guarantee as rollback consistency: a complete abort must restore the state the model attends, not just the transcript. The key failure is cross-layer: a correct logical rollback need not compose with retained inference state, and the gap can remain invisible to the application. To isolate cache effects from text effects, we introduce a same-token/different-cache audit that holds decision-step tokens identical while varying only whether the cached prefix is stale or rebuilt from committed state. Across seven open-weight families (3.8B-36B), retained KV alone flips a typed protected effect in 25 of 63 audited cells, while attacker tokens are absent from the served request in all 63; rebuilding the cache closes every cell. The channel reproduces in an end-to-end session application, on the default Hugging Face Transformers cache-reuse path, and under LangGraph time-travel, where verified logical rollback can still leave attended KV stale. Susceptibility varies across models, but the underlying attended-state integrity violation is structural. We rule out position and length confounds, generalize across protected effects, policy structures, and a cache-isolated Mixture-of-Experts model, and show that transaction-local cache restoration closes the channel without requiring a global cache flush. All headline results are deterministic and reproducible from released artifacts.

In expert-parallel (EP) MoE serving, every layer synchronizes at the slowest GPU. Dispatchers balance token counts (EPLB, LPLB, UltraEP) or activated-expert counts (METRO), assuming expert time is linear in one. Measurements on two datacenter GPU generations show it is neither: below $\nstar\!\approx\!156$--$168$ tokens, HBM weight streaming dominates---cost attaches to \emph{activated replicas}, not tokens; above it, grouped GEMM rounds tokens to 128-tile $M$-tiles, so \emph{splitting} an expert adds padded compute. A max-affine profile $t=\max(a+bG,\,c+βN)$ captures both regimes. Realistic decode batches hold hot experts in the linear regime and cold in the flat \emph{simultaneously}; recorded batches show proxy dispatches differ by $1.4$--$1.6\times$ in modeled block time (p95 up to $1.7\times$), and \emph{which} proxy wins flips with the regime. We formalize per-batch dispatch as a fixed-charge makespan problem---NP-hard on two fully replicated GPUs, polynomial in degenerate limits---and present \sys{}, a makespan-aware dispatcher solving it in milliseconds off the critical path; its SGLang integration runs out-of-process and fuses dispatch with count collection into one in-graph kernel. Anchored by an 8-GPU Testbed~A microbenchmark, \sys{} stays within 1\% of the best fixed baseline everywhere and wins by up to $15.5\%$ where regimes mix. End-to-end on Testbed~B, Qwen3-235B (inside the win region) gains $4$--$6\%$ throughput and cuts p99 latency by ${\sim}15.6\%$; DeepSeek-V3 (outside, communication-dominated) shows only mechanism cost. A phase diagram, not a universal win, is the claim: it predicts both outcomes before deployment.